[drtgh]

About six years ago I saw a tourist ordering in a restaurant by writing on his smartphone, translating, and using a classic text-to-speech function while showing the screen to the waitress, and it worked him pretty well for ordering. Since then the people I saw using it increased. Common sentences are used.

The problem would be in a conversation between two persons. Nowadays automated text translations are not reliable, can introduce even opposite meanings and they are not aware of nuances; it needs active supervision at this moment (and the following years).

With voice, a time-delay is needed for to acquire sentence context if the sources and target languages share structures, or a mandatory time-delay when the language structures are different, and also a general time-delay would be recommended for to avoid the interlocutors to listen two voices at same time. I'm not sure the real-time can be done like the ones we see in Star Trek (with voice to voice at least).

Important note: I would not recommend to popularize the synthesis of our personal voices. Variations from some reference models would be much better.